Mercury 3: Speed, Benchmarks, Pricing and Key Specs

Quick take: Inception, a startup that builds diffusion-based language models, has released Mercury 3. In its own benchmarks, the model beat OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 on eight of 12 tests while running four to eight times faster. No outside group has published results yet, so treat the numbers as claims.

A Different Way to Write Text

Most chatbots write one token at a time, left to right. Mercury works differently. A diffusion language model drafts the whole answer at once, then refines every position in parallel over a few passes. That is the idea behind its speed. The question hanging over the approach has been whether it can scale, and Inception says Mercury 3 is its answer.

Inception's Mercury 3 Claims It Beats GPT-6 Luna and Claude Haiku 5.5 at Up to 8x the Speed

Mercury 3 vs GPT-6 Luna vs Claude Haiku 5.5

Metric	Mercury 3	Claude Haiku 5.5	GPT-6 Luna
Speed (tokens per second, on Nvidia GPUs)	1,068	About 240	About 130
Speed advantage	Baseline	About 4.5x slower	About 8x slower
SWE-Bench Atlas (coding)	51.1	32.3	33.3
DeepSWE (coding)	46.3	51.6	Not reported
List price per million input tokens	$0.20	$0.10	$0.10
List price per million output tokens	$0.70	$0.50	$0.50
Launch price (input / output)	$0.10 / $0.35	N/A	N/A

Figures are from SiliconANGLE’s report and Inception’s own benchmarks. Speed ratios are our calculation from the tokens-per-second figures. Competitor prices are for typical prompts.

Where Mercury 3 Wins and Loses

Wins: Eight of 12 benchmarks. Its biggest lead over GPT-6 Luna is in coding, on SWE-Bench Atlas.
Loses: It trails Haiku 5.5 on DeepSWE, a second coding test, and on one long-context test.
Close calls: GPT-6 Luna won two benchmarks, each by less than a point.

Key Specs

Detail	Information
Released	October 19, 2026
Context window	1 million tokens (up from 260,000 in Mercury 2.5)
Size	Not disclosed, but over 40 times larger than Mercury 1
Cached input	90% cheaper than standard input
Availability	Inception’s platform, OpenRouter, Baseten and Vercel
Compatibility	Works with code written for OpenAI’s API with minor changes

Inception has also said Mercury 2 and Mercury 2.5 are no longer available to new users.

Who Is Using It?

Early customers care about delays that users can feel. Octen AI, a real-time search startup, says an answer taking five seconds feels broken on smart glasses, and reports a median generation time of 0.9 seconds in its tests. Tutoring app RevisionDojo says hints and marked answers now return in about a second.

The Business Behind It

Inception was founded in 2024 by computer science professors Stefano Ermon, Aditya Grover and Volodymyr Kuleshov. Last November, it raised $50 million in a round led by Menlo Ventures, with Microsoft’s M12 among the backers. Reuters reported in May that Microsoft was in talks to acquire it, with the startup seeking more than $1 billion, and that SpaceX had also approached it. No deal has been announced.

What to Take Away

The Case For	The Case For Caution
Much faster output, useful for voice, search and agents	Benchmarks were run by Inception, not an independent lab
1 million token context window	Costs more per token than both rivals at list price
Strong coding result on SWE-Bench Atlas	Loses on DeepSWE and one long-context test

If speed is your bottleneck, Mercury 3 deserves a test. If you care about price per token or consistent coding results, wait for independent benchmarks. Learn more about large language models, and see more AI coverage in our technology section on TechnoSports, including our report on Anthropic’s draft IPO filing.

Sources: SiliconANGLE, Business Wire, Inception, Unite.AI.